PaperlayerView exact revision on arXiv

Source-linked AI summary

Transfer Learning with intelligent training data selection for prediction of Alzheimer's Disease

Naimul Mefraz Khan, Marcia Hon, Nabila Abraham

arXiv:1906.01160v1cs.CVcs.LG

TL;DR

The paper addresses MRI-based Alzheimer’s diagnosis under deep learning’s dependence on large annotated datasets and careful optimization. It combines VGG transfer learning with layer-wise tuning and entropy-based slice selection, achieving state-of-the-art results across three classification tasks using training sets 10–20 times smaller than contemporary methods. The study also examines tuning depth and data selection on ADNI.

  • Problem

    Deep learning for MRI-based Alzheimer’s diagnosis depends on large annotated datasets and careful network optimization, while medical annotations can be expensive and privacy-limited.

  • Method

    The method initializes a VGG architecture with natural-image pretrained weights, progressively fine-tunes selected layer groups, and uses image entropy to select informative MRI slices.

  • Results

    10–20 times smaller training sets achieved state-of-the-art performance for AD vs NC, AD vs MCI, and MCI vs NC classification.

  • Takeaways & Limitations

    Transfer learning with intelligent slice selection can support Alzheimer’s classification while substantially reducing reliance on large training sets.

Abstract

from arXiv · show

Detection of Alzheimer's Disease (AD) from neuroimaging data such as MRI through machine learning has been a subject of intense research in recent years. Recent success of deep learning in computer vision has progressed such research further. However, common limitations with such algorithms are reliance on a large number of training images, and requirement of careful optimization of the architecture of deep networks. In this paper, we attempt solving these issues with transfer learning, where the state-of-the-art VGG architecture is initialized with pre-trained weights from large benchmark datasets consisting of natural images. The network is then fine-tuned with layer-wise tuning, where only a pre-defined group of layers are trained on MRI images. To shrink the training data size, we employ image entropy to select the most informative slices. Through experimentation on the ADNI dataset, we show that with training size of 10 to 20 times smaller than the other contemporary methods, we reach state-of-the-art performance in AD vs. NC, AD vs. MCI, and MCI vs. NC classification problems, with a 4% and a 7% increase in accuracy over the state-of-the-art for AD vs. MCI and MCI vs. NC, respectively. We also provide detailed analysis of the effect of the intelligent training data selection method, changing the training size, and changing the number of layers to be fine-tuned. Finally, we provide Class Activation Maps (CAM) that demonstrate how the proposed model focuses on discriminative image regions that are neuropathologically relevant, and can help the healthcare practitioner in interpreting the model's decision making process.

I. INTRODUCTION

The paper targets MRI-based Alzheimer’s diagnosis while addressing deep learning’s dependence on large annotated datasets and extensive optimization. It combines transfer learning, layer-wise tuning, and entropy-based slice selection to pursue state-of-the-art classification with substantially less training data.

  • I. INTRODUCTION: Deep learning methods require extensive annotated data, computational resources, and parameter tuning, all problematic for medical imaging.Physician-annotated data can be expensive and restricted by privacy and ethical concerns.
  • I. INTRODUCTION: Transfer learning reuses CNNs pretrained on large natural-image datasets and adapts them to new image representations using smaller datasets.This approach leverages established architectures and learned image features instead of training entirely from scratch.
  • I. INTRODUCTION: The method progressively retrains only selected upper CNN layers while keeping lower-level layers frozen.The authors report that only a few top layers are needed for the best result, reducing required training time.
  • I. INTRODUCTION: Image entropy selects informative MRI slices because random sampling may not represent sufficient structural variation.The selection strategy is intended to make a small training set more informative.
  • I. INTRODUCTION: 10–20 times smaller training sets still achieved state-of-the-art results across AD vs NC, AD vs MCI, and MCI vs NC classification.The approach combines transfer learning with intelligent training-data selection.
  • I. INTRODUCTION: The framework learns discriminative MRI regions from labeled slices and later categorizes individual slices as AD, MCI, or NC.The deployed model uses the learned distinctive representation for classification.

II. RELATED WORKS

Prior work progressed from classical MRI classifiers to deep learning, but practical diagnosis remains constrained by the need for large, costly annotated datasets. This paper addresses that gap with transfer learning and intelligent data selection.

  • II. RELATED WORKS: Classical methods such as SVMs and feed-forward neural networks have also been applied successfully to structural MRI diagnosis.One cited approach combines dual-tree complex wavelet features with a feed-forward neural network.
  • II. RELATED WORKS: Deep learning approaches for AD diagnosis include autoencoders, 2D and 3D convolutional networks, stacked autoencoders, and established CNN architectures.Most reported experiments use the ADNI benchmark database and binary classification tasks.
  • II. RELATED WORKS: The proposed method reports improved accuracy while using a training size 10–20 times smaller than methods in the contemporary literature.The stated motivation is to reduce dependence on physician-annotated data that may be expensive or unavailable.

A. CONVOLUTIONAL NEURAL NETWORKS AND TRANSFER LEARNING

CNNs learn local image features through convolution and progressively reduce representation size through pooling. Transfer learning addresses the data demands of training such networks from scratch by adapting pretrained models to new imaging tasks.

  • A. CONVOLUTIONAL NEURAL NETWORKS AND TRANSFER LEARNING: CNNs combine convolutional and pooling layers with fully connected and output layers to transform images into predicted labels.Convolution extracts local features, while max pooling reduces feature-map size by selecting local maxima.
  • A. CONVOLUTIONAL NEURAL NETWORKS AND TRANSFER LEARNING: Large CNN parameter counts make training from scratch dependent on many samples for backpropagation to converge properly.Limited training data can leave optimization trapped at a local minimum.
  • A. CONVOLUTIONAL NEURAL NETWORKS AND TRANSFER LEARNING: Fine-tuning pretrained CNNs has generally outperformed training from scratch in several medical-imaging applications.Prior uses include ultrasound-plane localization, interstitial-lung-disease classification, and cardiac-image retrieval.

A. NETWORK ARCHITECTURE

The model follows the deep VGG-19 architecture, adapting its final classifier for MRI-based Alzheimer’s diagnosis. Layer-wise transfer learning evaluates progressively freezing convolutional groups while retraining the remaining upper layers.

  • A. NETWORK ARCHITECTURE: VGG was selected for its accuracy, efficiency, and adaptability beyond ImageNet, including computer-aided diagnosis applications.Its design increases network depth while keeping other network parameters fixed.
  • A. NETWORK ARCHITECTURE: The network follows VGG-19 with 16 convolutional layers arranged in five blocks and a modified final classification layer.The architecture uses 3×3 convolution filters and five max-pooling layers.
  • A. NETWORK ARCHITECTURE: The four transfer-learning configurations freeze convolutional layers 1–4, 1–8, 1–12, or all 1–16 layers.These groups correspond closely to blocks in the VGG-style architecture.
  • A. NETWORK ARCHITECTURE: Freezing Block 1 separately from Block 2 changed average accuracy by only 0.05–0.45%, so the authors used grouped configurations to speed experimentation.The authors associate Block 1 with low-level feature extraction.

B. MOST INFORMATIVE TRAINING DATA SELECTION

The method selects MRI slices with the highest image entropy rather than sampling randomly, aiming to retain informative structural variation while using fewer training images.

  • The selection strategy replaces random training-image sampling, which may fail to represent sufficient structural variation in a small dataset.
  • Highest-entropy MRI slices are selected because entropy measures slice variation and higher entropy indicates more information.The images are preprocessed for noise removal and standardized before ranking slices by entropy.
  • Image entropy is calculated from an image slice’s intensity histogram before slices are sorted in descending order.

A. DATASET

Experiments use the ADNI benchmark dataset, comprising equal groups of AD, MCI, and NC subjects and supporting three binary classification tasks plus a three-way task.

  • ADNI contains 150 subjects: 50 each with Alzheimer’s Disease, Mild Cognitive Impairment, and Normal Control.
  • The study evaluates AD vs. NC, AD vs. MCI, and MCI vs. NC binary classification, with the latter two described as more difficult because MCI differences are subtle.
  • The experiments also include three-way classification to assess the proposed model’s robustness.

B. RESULTS FROM LAYER-WISE TRANSFER LEARNING

Layer-wise transfer learning was evaluated with entropy-selected datasets containing 8, 16, or 32 images per subject. The results show that the optimal number of trainable layers depends on training-set size, while entropy-based selection outperforms training from scratch and random selection.

  • Results from layer-wise transfer learning: The results support the view that early CNN layers learn broadly applicable low-level features, whereas later layers learn application-specific features.Thus, fine-tuning the last few layers is often sufficient, but the optimal choice depends on the application and training-set size.
  • Results from layer-wise transfer learning: With 8 images per subject, fine-tuning all layers gives the highest accuracy, although the very small training set may produce underfitting or overfitting.The 16-images-per-subject setting shows no clear trend in accuracy across layer-freezing groups.
  • Results from layer-wise transfer learning: With 32 images per subject, freezing all convolutional layers and fine-tuning only the fully connected layers is sufficient, consistent with the observed monotonic trend.This setting uses fewer parameters and therefore supports faster training.
  • Results from intelligent training data selection: The proposed entropy-selected transfer-learning model significantly outperforms training from scratch and random slice selection across all classification problems.The performance gap widens as the training set becomes smaller, particularly because random selection is unlikely to capture sufficient variation from only 8 slices per MRI volume.
  • Results from intelligent training data selection: Adding 64 images per subject to training from scratch provides little or no improvement and sometimes slightly decreases accuracy.The authors attribute this to lower-entropy slices adding redundant or noisy information after the most informative slices have been selected.

C. COMPARISON WITH EXISTING METHODS

The proposed method outperforms existing deep-learning approaches while using substantially fewer training images, and its CAM analysis highlights medically relevant MRI regions in correct predictions.

  • C. COMPARISON WITH EXISTING METHODS: 4% and 7% accuracy gains over state-of-the-art methods were achieved for AD vs. MCI and MCI vs. NC, respectively.The method used 32 images per subject and averaged a 4.5% accuracy increase over the state-of-the-art 3D CNN.
  • C. COMPARISON WITH EXISTING METHODS: Training data were reduced approximately 10 times, or almost 20 times with 16 images per subject, while retaining competitive accuracy.The 16-images-per-subject setting had a slight accuracy reduction but remained superior to most existing methods.
  • C. COMPARISON WITH EXISTING METHODS: The method also achieved state-of-the-art performance for three-way AD, MCI, and NC classification using the same architecture with a modified final layer.Evaluation used five-fold cross-validation with an 80%-20% training-testing split.
  • C. COMPARISON WITH EXISTING METHODS: CAM overlays visualize model attention on query images, with redder regions indicating higher attention and examples shown for both correct and incorrect diagnoses.The CAMs were generated from the three-way classification model.
  • C. COMPARISON WITH EXISTING METHODS: For a correct prediction, the network emphasized gray matter and cerebrospinal-fluid regions, aligning with Alzheimer’s-related atrophy and increased CSF.The authors interpret this alignment as neuropathologically relevant evidence for the model’s decision focus.

V. CONCLUSION

The paper proposes transfer learning with intelligent MRI-slice selection for Alzheimer’s diagnosis and reports strong comparative and three-way classification results. It identifies broader diagnostic applications and improved probabilistic image selection as future directions.

  • V. CONCLUSION: The proposed transfer learning method significantly outperformed six state-of-the-art methods, improving accuracy by 4% for AD vs. MCI and 7% for MCI vs. NC.Experiments used the ADNI dataset with 50 subjects from each of AD, MCI, and NC categories.
  • V. CONCLUSION: The method achieved state-of-the-art results for three-way classification, supporting its reported robustness.The paper also examined layer-wise transfer learning and its relationship to training-data size.
  • V. CONCLUSION: Future work will test the architecture on other computer-aided diagnosis problems and improve entropy-based image selection with additional probabilistic measures.These directions are presented as investigations rather than established capabilities.
Loading 1906.01160v1…